Learning Content:
Progress heavy Recognition (Automation Technology .2019) based on the depth study of pedestrian - Chapter 1-2
Key Technology of progressive heavy pedestrian recognition (PhD thesis .2019) - Chapter 1
[Because just getting started with Re-ID, yet the depth of learning to learn and understand more superficial, just reading the Liberal part of the algorithm to be the basis of completion of the chapter go to understand. ]
1. What is heavy pedestrian recognition?
Pedestrian weight identification (Person re-identi fi cation), also known as a pedestrian again identified, is widely regarded as the sub-problem of an image retrieval, whether there is a specific pedestrian technical computer vision technology to determine the image or video, i.e., given a monitor pedestrian image the pedestrian image retrieval across the device. pedestrian recognition weight limitations can make up for the visual fixing the camera, and detecting a pedestrian, the pedestrian tracking technology, used in video surveillance, intelligent security, and other fields.
Popular understanding of: tracking the same person in a different camera.
2, re-constituting the pedestrian recognition system: heavy pedestrian detection Pedestrian Recognition +
3, the difficulty of identification:
(1) similar to the appearance of a pedestrian different;
(2) the same pedestrian different postures at different angles;
(3) a pedestrian is blocked;
Differences in light conditions (4) to collect data.
4, based on the weight of the pedestrian image recognition:
(1) Based on the learning process is characterized:
Pedestrian image is divided into a plurality of regions, the underlying design of a variety of different local visual features (such as color, texture, shape, etc.) or middle feature (e.g., length of hair, the color of shoes, belongings and the like) for each region, and from the global they can be combined to form an angle to reflect characteristics represented pedestrian appearance attributes.
(2) based on distance metric learning method:
Focus on learning to identify different cameras distance metric or a pedestrian environment discrimination subspace, so that the distance between the same pedestrian is sufficiently small, the distance between the different pedestrian sufficiently large.
(3) Based on the depth of learning methods:
The characteristics of a learning and a measure of learning two separate aspects of joint depth in a unified neural network model.
Application of related technologies:
GAN: generating confrontation Network
CNN: convolution neural network